Unveiling soil-vegetation interactions: Reflection relationships and an attention-based deep learning approach for carbon estimation

Estimating soil organic carbon (SOC) from satellite imagery, particularly in areas with both bare soil and vegetation, poses significant challenges. Traditional approaches often overlook the complex interactions between soil and vegetation. Addressing this gap, our study introduces an innovative method that leverages novel correction of hyperspectral reflections to adjust for vegetation levels, enhancing SOC estimation accuracy. Moreover, we propose an attention-based deep neural network that dynamically prioritizes spectral features crucial for SOC prediction. This mechanism significantly improves the model's ability to detect significant features for accurate SOC estimation. Comparative experiments with traditional models on a benchmark dataset demonstrate our method's effectiveness in reducing vegetation influence and accurately estimating SOC across mixed landscapes. Our findings represent a notable advancement in SOC estimation from satellite imagery, highlighting the potential of advanced learning-based techniques with attention-driven feature weighting for SOC estimation.